A cross-mining area spectral mineral identification optimization method based on federated learning

Abstract

The application discloses a cross-mining area spectral mineral identification optimization method based on federated learning, comprising: a central server initializes a global model, each mining area is used as a client to pre-train and continuously learn locally using private data, and only encrypted model updates are uploaded; the server securely aggregates updates through a weighting algorithm, generates and distributes a new generation of global model; under the premise of strictly protecting data privacy, the method effectively aggregates multi-source knowledge, trains a mineral identification model with strong generalization ability, and can continuously absorb new data knowledge through a continuous learning mechanism, thereby significantly improving identification accuracy and robustness in a complex geological environment.

Description

Technical Field

[0001] This invention relates to the field of spectral mineral identification technology, and in particular to an optimized method for cross-mining area spectral mineral identification based on federated learning. Background Technology

[0002] In the current field of mineral resource exploration, hyperspectral remote sensing technology has become a key means of identifying surface and underground minerals. However, the practical application of this technology faces two major bottlenecks. First, valuable mineral spectral data and precise annotations are scattered among various mining companies and research institutions, forming data silos. Due to strict restrictions on trade secrets, data security, and ownership, this private data containing core geological knowledge cannot be centrally shared, making it difficult to implement traditional centralized machine learning methods and unable to aggregate knowledge from multiple sources to train models with strong generalization capabilities.

[0003] Secondly, exploration itself is a continuous and dynamic process, especially in underground drilling scenarios. After borehole washing, the borehole wall surface is rough and uneven and is constantly extending, generating new data continuously. Existing methods mostly perform one-time modeling on static historical datasets, which cannot enable the model to continuously and securely learn and evolve from new data while protecting the privacy of historical data from all parties. This results in outdated model knowledge, making it difficult to adapt to new blocks or newly discovered geological phenomena. Summary of the Invention

[0004] This application provides a federated learning-based optimization method for cross-mining area spectral mineral identification, which improves the accuracy, robustness, and adaptability to unknown geological environments of mineral identification.

[0005] This application provides an optimized method for cross-mining area spectral mineral identification based on federated learning, including: S1. A global mineral identification model with a neural network architecture for processing hyperspectral data is built in the central server. Each participating mining area, as a client, deploys the same model architecture locally and prepares its own private historical spectral dataset. S2, each client uses a private historical spectral dataset and few-shot augmentation technology to pre-train a model locally and inject domain prior knowledge; S3, each client encrypts the gradient update of the trained model and uploads it to the central server. The server uses a weighted algorithm designed for non-independent and identically distributed models to perform secure aggregation and generate a new generation of global models. S4, each client establishes a local knowledge consolidation library to store key knowledge representations and continuously monitors for newly added local exploration data; S5: When the new data reaches the threshold, the client automatically triggers local training, and combines the new data with the playback of the knowledge consolidation base to update the model using a continuous learning algorithm. In S6, the client uploads the updated gradients in encrypted form, and the server performs secure aggregation again to generate and distribute the next-generation global model after iteration.

[0006] Preferably, the private historical spectral dataset specifically includes: local assets independently owned and controlled by each participating mining area, whose ownership, management rights, and access rights belong to a single mining area entity and are not shared with other participants or a central server; the dataset contains spectral information and corresponding mineral category labels that have been obtained and labeled through historical exploration activities before the start of federated collaborative training; the spectral information in the dataset covers continuous, narrow-band reflectance or emissivity information from visible light to thermal infrared bands; the mineral category labels correspond to each sample in the spectral information and are determined in advance through geological verification methods; the private historical spectral datasets of each participating mining area have significant differences in the spectral characteristics, co-occurrence combinations, and background environment of mineral composition, do not satisfy the independent and identically distributed hypothesis, and there are cases of scarcity of specific mineral category samples and imbalance in the number of samples of different categories.

[0007] Preferably, the next-generation global model includes: a central server collecting privacy-processed model update gradients uploaded by each client; aggregating the collected model update gradients using a weighted algorithm to generate an aggregated model update; applying the aggregated model update to the global model parameters of the previous round to generate a next-generation global model; and the central server storing and managing the version of the next-generation global model and distributing it to each client.

[0008] Preferably, the continuous learning algorithm for model updating specifically includes: the client continuously monitoring newly added exploration data locally, and automatically triggering local training when the accumulated data reaches a preset threshold; during training, the newly added data is mixed with historical core knowledge playback data extracted from the local knowledge consolidation base to form a training set; the local model is trained using a continuous learning algorithm, which prevents the model from forgetting consolidated historical knowledge when learning new data by introducing regularization constraints or directly replaying historical data; after training is completed, the local knowledge consolidation base is updated, and the model update gradient is calculated.

[0009] Preferably, the continuous monitoring of newly added local exploration data specifically includes: S41, identifying hyperspectral scan data from the borehole wall from the newly added exploration data; S42, performing local synchronous registration of the hyperspectral image and the 3D point cloud for the newly added borehole wall hyperspectral data, and assigning each spectral pixel its corresponding 3D spatial attribute; S43, on the client side, using a physically inspired neural network module, combining 3D geometric features to perform roughness-aware correction on the original borehole wall spectrum, generating a spectrum equivalent to that under standard observation conditions; S44, adopting a spectral-spatial dual-stream network architecture, extracting the deep features of the corrected spectrum and the spatial context features of the 3D point cloud in parallel on the client side, and performing adaptive fusion to complete mineral identification; S45, generating enhanced samples from the corrected high-quality data and identification results, and extracting key anti-interference knowledge from them and storing them in a local knowledge consolidation base; wherein, the 3D spatial attributes include surface normal vector and local curvature, used to quantitatively describe the roughness and micro-topography of the borehole wall after cleaning.

[0010] Preferably, the generation of a spectrum equivalent to that under standard observation conditions includes: locally on the client side, combining the original pore wall spectrum with its three-dimensional geometric features and inputting it into a physics-inspired neural network correction module; the network learns the mapping relationship from actual rough observation conditions to standard flat observation conditions by integrating prior physical laws of the interaction between light and rough surfaces; the network outputs a spectral correction value and uses this correction value to process the original pore wall spectrum; after processing, a corrected spectrum is generated, which is equivalent to the spectrum of the target mineral measured under standard laboratory observation conditions.

[0011] Preferably, the mineral identification process includes: employing a spectral-spatial dual-stream network architecture to process the spectral information and three-dimensional spatial information in parallel on the client side; extracting deep spectral features and spatial context features through the dual-stream network; dynamically fusing the spectral features and spatial features through an adaptive fusion module in the multi-layer structure of the network; inputting the features generated after the final hierarchical fusion into a classifier, and outputting the mineral category and its confidence level assessment corresponding to each spatial unit.

[0012] Preferably, the step of identifying hyperspectral scanning data from the borehole wall in the newly added exploration data specifically involves: S411, after identifying the borehole wall hyperspectral data, simultaneously checking whether it carries complete inertial measurement unit motion data, and automatically triggering or skipping the motion blur correction process accordingly; S412, parsing the motion data locally on the client side, calculating and reconstructing the point spread function that causes spectral image degradation in conjunction with the scanning model; S413, using a specially designed motion-invariant feature network, taking the original blurred spectrum and point spread function as joint inputs, directly extracting deep-level sharpening spectral features robust to motion interference locally; S414, converting the sharpening features output by the network into enhanced hyperspectral image data.

[0013] Preferably, the automatic triggering or skipping of the motion blur correction process includes: the client checking whether the newly added aperture wall hyperspectral data carries complete inertial measurement unit motion data, and performing integrity checks on the motion data for timestamp alignment, data dimension completeness, and sampling rate coverage; if the motion data passes the integrity check, the subsequent motion blur correction process is automatically triggered; if the motion data does not carry the complete motion data or the check fails, the motion blur correction process is skipped.

[0014] Preferably, the deep-clarification spectral features specifically include: inputting the blurred spectral information and the corresponding point spread function as joint inputs into a motion-invariant feature network; the network reverses the blurring process in the feature space by fusing the physical degradation prior encoded by the point spread function; the network focuses on extracting essential spectral-spatial patterns that are robust to motion disturbances and related to mineral categories through an invariant feature extraction layer; and finally outputs a deep spectral-spatial feature map, which is semantically equivalent to the features extracted from a still, clear image.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages: By employing a federated learning framework, and ensuring that the original spectral data from each mining area remains local while strictly protecting data privacy and ownership, the model successfully aggregates differentiated knowledge scattered across different mining areas, collaboratively training a global mineral identification model with stronger generalization capabilities. This model not only effectively overcomes the limitations of limited and unevenly distributed data samples from single mining areas, but also, through an integrated continuous learning mechanism, continuously absorbs new data knowledge during dynamic exploration while avoiding forgetting historical experience. This achieves continuous evolution and lifelong learning of model performance, ultimately significantly improving the accuracy, robustness, and adaptability to unknown geological environments in mineral identification.

[0016] By employing localized spectral correction and dual-stream fusion identification to address borehole wall roughness and micro-shadowing interference, the accuracy of mineral identification in complex field environments is significantly improved. The entire process, from 3D registration and physically-inspired correction to feature fusion, is completed locally on the client side, strictly ensuring the privacy and security of the original exploration data. This method effectively eliminates the interference of surface morphology on the spectrum, enabling the model to learn more essential mineral features. This not only significantly enhances the accuracy and robustness of single-point identification but also provides high-quality nutrients for the continuous learning and evolution of the federated system through its high-quality correction samples and anti-interference knowledge, thereby improving the overall practicality and environmental adaptability of the collaborative model.

[0017] By integrating motion sensing and correction capabilities locally on the client side, the problem of hyperspectral data blurring and degradation caused by the movement and vibration of downhole probes is effectively solved, thus significantly improving the accuracy of mineral identification. Under the premise of strictly protecting the data privacy of each mining area, this solution enables the federated learning system to extract robust features from higher-quality, clearer borehole wall spectral data. This not only enhances the adaptability and generalization ability of the global model in complex dynamic acquisition environments but also lays a solid data foundation for achieving accurate and reliable cross-mining area collaborative identification. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an optimized cross-mining area spectral mineral identification method based on federated learning, according to an embodiment of the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terminology used herein includes any and all combinations of one or more of the associated listed items.

[0021] Example 1: Figure 1 This is a flowchart illustrating an optimized cross-mining area spectral mineral identification method based on federated learning, according to an embodiment of the present invention.

[0022] like Figure 1 As shown, a cross-mining area spectral mineral identification optimization method based on federated learning includes the following steps: S1. A global mineral identification model with a neural network architecture for processing hyperspectral data is built in the central server. Each participating mining area, as a client, deploys the same model architecture locally and prepares its own private historical spectral dataset.

[0023] In this context, private historical spectral datasets specifically refer to local data assets that are independently owned and controlled by each participating mining area (client) and are not shared with other participants or the central server. The specific connotations include: ownership and privacy. "Private" means that the ownership, management, and access rights of the data belong entirely to a single mining area entity. "Historical" refers to data accumulated and labeled by each mining area through previous exploration activities (such as aerial / ground hyperspectral remote sensing, borehole core scanning, etc.) before the start of federated collaborative training, distinguishing it from newly generated data after the start of training.

[0024] The dataset primarily consists of paired hyperspectral data samples and their annotations. Hyperspectral data is typically image cubes or spectral curves, containing reflectance or emissivity information for ground features across multiple continuous, narrow bands from visible light to thermal infrared. The annotations are mineral category labels (e.g., calcite, muscovite, limonite) corresponding to each sample (pixel or region) in the hyperspectral data. These labels are pre-determined by experts in the mining area through field verification and laboratory analysis.

[0025] The core characteristics are: Non-independent and identically distributed (i.i.d.) data distribution (i.e., spectral characteristics, co-occurrence assemblages, and background environment of specific minerals) varies significantly across mining areas due to differences in geographical location, geological background, and mineralization type, failing to meet the independent and identically distributed assumption required by traditional machine learning. Small sample size and imbalance mean that the labeled sample size for certain minerals in a single mining area may be very small (small sample size), and the sample size for different mineral categories may vary greatly (imbalance). Value carrier: These datasets represent the core knowledge assets of each participant, containing their unique geological insights. The purpose of federated learning is precisely to distill and aggregate this knowledge by exchanging model updates, without aggregating these data ontology.

[0026] Specifically, on the central server, a neural network model optimized for processing hyperspectral data, namely a global mineral identification model, is designed and built. This model architecture must fully consider the characteristics of spectral data, such as effectively extracting spectral features and being compatible with multi-band (e.g., VNIR-SWIR, MWIR, LWIR) data input and feature fusion. The model's parameters are initialized (usually randomly), generating an initial version (V0) of the model.

[0027] Each participating mining area acts as a client node, deploying a mineral identification model locally (on its internal server or computing platform) with the exact same architecture as the central server. Each client pulls (downloads) the initialized global model parameters (V0) from the central server and uses them to initialize its local model copy. At this point, all client local models have completely identical initial parameters.

[0028] Each client prepares its own private historical spectral dataset locally. The client can perform necessary preprocessing on the local dataset, such as noise removal, standardization, and partitioning into training and validation sets.

[0029] S2: Each client uses its private historical spectral dataset and few-shot augmentation techniques to pre-train models locally, injecting domain prior knowledge.

[0030] Specifically, each client loads its private historical spectral dataset, performs preprocessing operations such as noise filtering, spectral standardization / normalization, and bad band removal on the data, and divides the dataset into a training set for local pre-training and a validation set for evaluating the pre-training effect (or retains some data as a training set for subsequent federated training).

[0031] To address the issues of insufficient sample size and class imbalance in local data, the client employs small-sample augmentation techniques to expand and enrich the training samples. Specific techniques may include (but are not limited to): spectral feature matching and pseudo-label generation, comparing the spectral curves of unlabeled or scarce categories with features (such as absorption valley location, depth, and symmetry) from a standard mineral spectral library. Pseudo-labels are automatically assigned to samples with high matching scores, effectively increasing the number of training samples for specific scarce mineral categories. Spectral data synthesis and transformation involves adding noise, interpolating bands, and performing spectral mixing simulations on existing sample spectra to generate new synthetic samples, increasing data diversity. After this stage, the client obtains an augmented local training set with better sample size and class balance than the original dataset.

[0032] The client uses an enhanced local training set to train the model downloaded and initialized from the central server (i.e., the copy of the model after S1). The training objective is to minimize the model's prediction error on local data (such as cross-entropy loss). This process allows the model parameters to adapt to the client's local data distribution and mineral spectral characteristics, that is, to encode domain prior knowledge into the model weights.

[0033] After training is complete, the client evaluates the performance of the pre-trained model on a local validation set to confirm its learning effectiveness. The pre-trained model parameters and states (or the calculated preliminary model update gradients) are saved locally as the starting point for participating in subsequent formal federated learning rounds.

[0034] In S3, each client encrypts the gradient updates of the trained model and uploads them to the central server. The server uses a weighted algorithm designed for non-independent and identically distributed models to perform secure aggregation and generate a new generation of global models.

[0035] Specifically, each participating mining area (client) trains its current local model using its private historical spectral dataset (or an augmented dataset). After training, the client calculates the difference between its local model parameters and the previous round's global model parameters obtained from the central server; this difference is the model update gradient. The core principle is: only gradients are calculated; raw private data is never shared or uploaded.

[0036] To ensure privacy and security, and to prevent the original data from being reverse-engineered during transmission or on the server side, each client must encrypt or perturb the model update gradients before sending them out locally. Common techniques include: homomorphic encryption, which encrypts the gradients, allowing the server to directly aggregate and calculate the encrypted gradients in ciphertext without decryption, providing high-strength security; and differential privacy, which adds carefully designed random noise that meets differential privacy requirements to the gradients. This protects the privacy of any individual data point while ensuring the availability of the aggregated statistical results (i.e., the global model update). Each client uploads the encrypted or noisy model update gradients to the central server via a secure network channel. At this point, each participant has completed its contribution to this round of federated training.

[0037] After the central server collects the privacy-protected model update gradients uploaded by all participating clients, it performs an aggregation operation. To address the issues of non-independent and identically distributed (i.I.D.) data from different clients and imbalanced sample sizes, a weighted average is used instead of a simple average during aggregation. Weighting strategies aim to fairly and effectively reflect the contributions of each client. Common designs include: weighting by data volume, allocating weights based on the number of local samples used by each client for training in the current round. Clients with larger data volumes have higher weights in the aggregation. Weighting by update quality, dynamically allocating weights based on the client's training performance in this round (e.g., the decrease in training loss, improvement in local validation set accuracy, etc.) to encourage participants providing higher-quality updates. Objective-driven weighting, employing more advanced federated optimization algorithms (such as FedProx), introduces a regularization term during aggregation to constrain local updates from deviating too far from the global model, thereby improving aggregation stability and the generalization performance of the final model under non-independent and identically distributed data.

[0038] The central server applies the weighted and aggregated updates to the global model parameters from the previous round, thereby generating a new generation of global model. For example, if the previous round was version V1, this round generates version V2. The server stores and manages the version of this new global model parameters, preparing to distribute them to clients to begin the next round of training or for final application.

[0039] S4: Each client establishes a local knowledge consolidation library to store key knowledge representations and continuously monitors for newly added exploration data locally.

[0040] Among them, the local knowledge consolidation base is a set of highly representative sample feature vectors (high-level features (embedded vectors) extracted from the private historical spectral data of the mining area, which are key samples selected from historical data that can represent different mineral categories and are extracted after model forward propagation) or synthetic samples / features: pseudo samples or features that represent learned categories and are generated based on the distribution of historical data through generation technology. Key knowledge representation refers to abstract information extracted from the client's private historical spectral data that can effectively represent and distinguish the essential spectral characteristics of different mineral categories. It is the basic unit that constitutes the content stored in the local knowledge consolidation base.

[0041] S5: When the amount of new data reaches the threshold, the client automatically triggers local training. Combining the new data with the playback of the knowledge consolidation base, the model is updated using a continuous learning algorithm.

[0042] Specifically, the client background process continuously monitors the accumulation of newly added exploration data (such as newly acquired hyperspectral scan data). When the amount of new data or the novelty shown by the model's preliminary inference (for example, the model's prediction confidence for a large amount of new data is generally low, indicating that it may contain new patterns that the model has not fully learned) reaches a preset threshold, the local training task is automatically triggered.

[0043] The client prepares training data, which consists of a mixture of two parts: new data, newly acquired hyperspectral data (and its annotations or pseudo-annotations generated automatically) that triggers this round of training; and replay data, which extracts stored key knowledge representations or training samples reconstructed / represented by them from the local knowledge consolidation repository according to a certain strategy (such as uniform sampling or importance-based sampling). This part of the data represents the historical core knowledge that needs to be consolidated.

[0044] The client loads the latest local model (i.e., the global model obtained after the previous round of federated aggregation, or the model updated after the previous round of local continuous learning). This model is trained using the constructed hybrid dataset. The training process employs a continuous learning algorithm, the core mechanism of which is to prevent catastrophic forgetting. Key methods include regularization-based approaches: adding a regularization term to the training loss function (e.g., Elastic Weight Consolidation - EWC). This regularization term anchors the model's memory of historically important parameters (typically corresponding to knowledge learned from the knowledge consolidation base samples), constraining these parameters from drastically changing when training on new data. Another approach is replay-based methods, directly inputting replay data from the knowledge consolidation base along with the new data into the model for training. By allowing the model to learn from both old and new data simultaneously, it is forced to consider both old and new knowledge during parameter updates.

[0045] After training, an updated local model is obtained, which has absorbed new knowledge and retained old knowledge. The client evaluates the new data and adds representative new knowledge that may be lacking in the current knowledge consolidation base (e.g., feature vectors of representative new samples selected through model uncertainty) to the local knowledge consolidation base. At the same time, some old representations may be eliminated according to the storage strategy, so as to realize the dynamic updating of the knowledge base.

[0046] The client calculates the parameter differences between the updated local model and the model before training, after this round of continuous learning training; this is the model update gradient for this round of local training. This gradient encapsulates the net knowledge changes learned by the model from new data and consolidation replays.

[0047] In S6, the client uploads the updated gradients in encrypted form, and the server performs secure aggregation again to generate and distribute the next-generation global model after iteration.

[0048] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By employing a federated learning framework, and ensuring that the original spectral data from each mining area remains local while strictly protecting data privacy and ownership, the model successfully aggregates differentiated knowledge scattered across different mining areas, collaboratively training a global mineral identification model with stronger generalization capabilities. This model not only effectively overcomes the limitations of limited and unevenly distributed data samples from single mining areas, but also, through an integrated continuous learning mechanism, continuously absorbs new data knowledge during dynamic exploration while avoiding forgetting historical experience. This achieves continuous evolution and lifelong learning of model performance, ultimately significantly improving the accuracy, robustness, and adaptability to unknown geological environments in mineral identification.

[0049] Example 2: In Example 1, the system continuously monitors newly added local exploration data and triggers local continuous learning when the data volume or novelty reaches a threshold to update the model and consolidate knowledge. However, this method assumes that all new data are under ideal or uniform observation conditions, failing to fully consider the unique challenges faced by newly added borehole wall hyperspectral data in actual drilling exploration—namely, the significant roughness of the borehole wall surface after washing and the shadows caused by micro-topography, which can lead to nonlinear and spatially heterogeneous distortions in the spectral reflectance intensity of the same mineral. If such severely disturbed raw data is mixed indiscriminately with relatively ideal data and uniformly input into the model for training, it is equivalent to making the model learn both pure and distorted spectral features simultaneously. This inevitably leads to inherent conflicts in feature extraction and generalization, making it difficult to make stable and accurate identifications under complex surface conditions. Since there are huge differences in roughness and shadow patterns in different mining areas, different borehole sections, and even different areas of the same borehole wall, this interference is not fixed noise but a complex function strongly correlated with local geometry. Using a uniform processing flow and model architecture has limited adaptability.

[0050] In some embodiments, continuously monitoring newly added exploration data locally, step S4 includes: S41 identifies hyperspectral scanning data from the borehole wall in newly added exploration data.

[0051] S42, for the newly added pore wall hyperspectral data, perform local synchronous registration of hyperspectral images and 3D point clouds, and assign each spectral pixel its corresponding 3D spatial attributes.

[0052] Among them, the three-dimensional spatial attributes include the surface normal vector and local curvature, which are used to quantitatively describe the roughness and micro-topography of the hole wall after hole washing.

[0053] S43, on the client side, utilizes a physics-inspired neural network module to perform roughness-aware correction on the original pore wall spectrum by combining three-dimensional geometric features, generating a spectrum equivalent to that under standard observation conditions.

[0054] Specifically, the client obtains the original aperture wall spectral vector and its corresponding three-dimensional geometric feature vector (mainly the local surface normal vector and curvature). Simultaneously, it loads a locally pre-built, physically inspired spectral correction module. This module is a lightweight neural network whose structural design incorporates prior knowledge of the fundamental physical laws governing the interaction between light and rough surfaces.

[0055] The original spectral vector and the 3D geometric feature vector are concatenated along the feature dimension to form a combined input feature. This combined feature is then input into a physically inspired spectral correction module. The network performs nonlinear transformations through multiple hidden layers to learn the complex mapping relationship from observation conditions on rough surfaces to observation conditions on standard flat surfaces.

[0056] During the forward propagation of the network, the design of specific network layers or loss functions may reflect physical constraints. For example, the network may implicitly learn and output a correction coefficient matrix or spectral distortion field related to the incident-observation geometry (derived from the normal vector). Ultimately, the network outputs a spectral correction (which may be an additive term, multiplicative coefficients, or a complete spectral transform).

[0057] The client applies the spectral correction values ​​output by the network to the original pore wall spectrum. Specific correction operations (such as band-by-band multiplication and addition or more complex transformations) are determined by the mapping learned by the network. After these operations, the corrected spectrum is obtained. This spectrum is theoretically equivalent to the spectrum of the same mineral measured under standard laboratory conditions (i.e., flat surface, uniform vertical illumination and observation), thus eliminating the influence of roughness and microshading.

[0058] The generated corrected standard equivalent spectrum, corresponding 3D coordinates (or pixel index), and original spectrum identifier are cached or output locally on the client and passed to subsequent step S44 for feature extraction and recognition. All data, including original data, 3D data, intermediate corrections, and final results, are processed and stored locally on this client, with no data transmitted externally.

[0059] S44 employs a spectral-spatial dual-stream network architecture, extracting deep features of the corrected spectrum and spatial context features of the 3D point cloud locally and in parallel on the client side, and then performing adaptive fusion to complete mineral identification.

[0060] Specifically, the client loads the corrected standard equivalent spectral data from the output of step S43, along with the registered 3D point cloud data. The client also loads the spectral-spatial dual-stream recognition sub-network from the federated global model, dedicated to local processing. This network comprises parallel spectral feature extraction and spatial context feature extraction streams.

[0061] The corrected spectral data is input into the spectral stream. This stream is typically composed of a convolutional neural network or a spectral attention module, which extracts and abstracts deep discriminative information such as absorption features and shape features from the spectrum layer by layer, outputting a high-dimensional spectral feature vector. Three-dimensional point cloud data (or local geometric feature meshes derived from it, such as normal maps, curvature maps, and elevation change maps) is input into the spatial stream. This stream may employ point cloud neural networks or three-dimensional convolution to extract spatial context feature vectors describing the microscopic topography, undulation patterns, and local structural relationships of the borehole wall surface after cleaning.

[0062] The two-stream network does not simply concatenate features at the end. Instead, adaptive fusion modules (such as attention-based fusion gates) are designed in multiple intermediate layers of the network. At each fusion point, this module dynamically calculates the complementary weights of spectral and spatial features, using spatial context information (such as whether the current pixel is in a shadow or convex region) as a guide to calibrate or enhance the spectral features, achieving fine-grained feature interaction and fusion, and generating a series of fused feature maps.

[0063] The fused features from the final layer are input into a classifier (usually a fully connected layer), which outputs the probability distribution of each pixel belonging to each mineral category, thus completing mineral identification. Simultaneously, the network can output an estimate of identification uncertainty. This value, which can be based on the entropy of the classification probability or calculated through a specific branch, is used to quantify the confidence level of the model's judgment under the current rough surface conditions.

[0064] The final mineral identification results (category labels), confidence scores, and uncertainty estimates are stored and output locally. Simultaneously, representative spectral feature vectors, spatial context feature vectors, and fused features from key layers extracted from the intermediate layers of the network are cached.

[0065] S45 generates enhanced samples from the corrected high-quality data and recognition results, and extracts the key anti-interference knowledge from them and stores it in the local knowledge consolidation library.

[0066] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By employing localized spectral correction and dual-stream fusion identification to address borehole wall roughness and micro-shadowing interference, the accuracy of mineral identification in complex field environments is significantly improved. The entire process, from 3D registration and physically-inspired correction to feature fusion, is completed locally on the client side, strictly ensuring the privacy and security of the original exploration data. This method effectively eliminates the interference of surface morphology on the spectrum, enabling the model to learn more essential mineral features. This not only significantly enhances the accuracy and robustness of single-point identification but also provides high-quality nutrients for the continuous learning and evolution of the federated system through its high-quality correction samples and anti-interference knowledge, thereby improving the overall practicality and environmental adaptability of the collaborative model.

[0067] Example 3: Addressing the roughness and microscopic shadows of the borehole wall after cleaning, three-dimensional geometric correction and dual-flow feature fusion were performed locally on the client side, significantly improving mineral identification accuracy on complex surfaces. However, this method relies on a crucial assumption: the acquired hyperspectral image itself is clear, and the spectral information within its pixels is accurate and spatially aligned. In actual downhole operations, mechanical vibration and instantaneous velocity changes are unavoidable as the scanning probe moves along the borehole wall. This leads to motion blur in the spatial dimension and inter-band information misalignment in the spectral dimension of the acquired hyperspectral image. While the correction process in Example 2 can handle static surface geometric distortion, it cannot solve this complex blurring problem related to the motion trajectory, introduced by the dynamic degradation during acquisition. Directly inputting such severely motion-blurred raw data into the subsequent geometric registration and feature fusion modules will result in increased registration errors and geometric correction distortion due to low-quality basic data. Ultimately, the entire fine processing flow is built on distorted data, severely limiting the improvement of recognition performance.

[0068] In some embodiments, in step S41, hyperspectral scanning data from the borehole wall are identified in the newly added exploration data, including: S411, after identifying the hyperspectral data of the pore wall, simultaneously checks whether it carries complete motion data of the inertial measurement unit, and automatically triggers or skips the motion fuzz correction process accordingly.

[0069] Specifically, the client parses the data packets of hyperspectral scanning data from the borehole wall, checks whether they explicitly state that synchronously acquired inertial measurement unit motion data are included, and locates the relevant data files or data streams.

[0070] If IMU data is included, the client loads and quickly verifies the integrity of the motion data. Verification criteria typically include: timestamp alignment—whether the time series of the IMU data has a clear and complete correspondence with the timestamps of each scan line of the hyperspectral image; data dimensional completeness—whether the IMU data contains the triaxial angular velocities and triaxial accelerations (or equivalent motion parameters) necessary to describe three-dimensional motion; and sampling rate and coverage—whether the sampling frequency of the IMU data is high enough to completely cover the probe motion details throughout the entire scan, without large areas of missing data.

[0071] Case A (Data integrity, trigger correction): If the IMU data passes the integrity check, the client automatically triggers the complete motion blur correction process of Example 3. The current data is marked as requiring motion correction, and the scan line timestamp sequence of the hyperspectral image is extracted. The original hyperspectral data, IMU data, and scan line timestamp sequence are then sent to the subsequent step S412 for processing.

[0072] Case B (Data missing or incomplete, skip correction): If the data packet does not include IMU data, or IMU data verification fails, the client determines that the current borehole wall data does not meet the motion correction conditions. The subsequent motion correction steps in Example 3 will be automatically skipped, and the original hyperspectral data will be directly marked as stable acquisition data. It will then be routed to step S42 of Example 2, i.e., the registration process between the hyperspectral image and the 3D point cloud will begin directly.

[0073] The client logs the decision locally, including data identifier, integrity verification result, and final route destination.

[0074] S412 parses motion data locally on the client side, and calculates and reconstructs the point spread function that causes spectral image degradation by combining the scanning model.

[0075] Specifically, the client loads the verified and complete raw IMU motion data (high-frequency sampled triaxial angular velocity and triaxial acceleration sequences) and the scan line timestamp sequence of the hyperspectral image, which were passed from step S411. Necessary preprocessing is performed on the IMU data, such as removing bias and performing coordinate transformation (from the IMU coordinate system to the probe optical coordinate system). The time axis of the IMU data is precisely aligned with the start integration timestamp of each line of the hyperspectral image to ensure that each line of scan data can be associated with a corresponding motion time sequence.

[0076] Numerical integration is performed on the IMU motion data (angular velocity, linear acceleration) corresponding to each scan line. Combined with the probe's initial pose (usually assumed to be relatively stationary or known at the start of the scan), the instantaneous motion trajectory of the probe during the entire exposure period of that scan line is reconstructed after integration. This trajectory describes the positional and orientation (angular) changes of the probe's optical center in three-dimensional space.

[0077] The client loads the probe's preset scanning model parameters, including but not limited to: the instantaneous field of view of the optical system, the angular resolution of a single pixel, the scanning direction (e.g., axial sweeping along the aperture wall), and the inter-row time interval. The reconstructed instantaneous motion trajectory of the probe is mapped and calculated using the optical-geometric scanning model. The model analyzes how the image of a fixed point on the ground (i.e., a tiny mineral point on the aperture wall) moves across the detector pixel array during the exposure time due to the probe's movement and rotation. This movement path (a two-dimensional curve covering several pixels in the spatial dimension and several bands in the spectral dimension) is quantized and calculated. This describes how the light energy of a stationary point source is dispersed across multiple pixels and multiple spectral channels.

[0078] The energy travel path calculated in the previous step is discretized and normalized to generate a two-dimensional numerical kernel. This kernel is the point spread function (PSF). The PSF is typically a small two-dimensional matrix. One dimension corresponds to the spatial direction (along the scan line direction or perpendicular to the scan line), and the other dimension corresponds to the spectral direction (different bands). Its value represents the weighted distribution of energy leakage from the real point to adjacent pixels and bands. The client generates one or more corresponding PSF kernels for each scan line (or a more granular time slice based on the rate of motion change). All the calculated PSF kernels are associated with their corresponding hyperspectral image row indices (or pixel block indices) and cached locally on the client. Finally, the original blurred hyperspectral data and the generated PSF kernel set are packaged and prepared for transfer to step S413.

[0079] S413 utilizes a specially designed motion-invariant feature network, taking the original blurred spectrum and point spread function as joint inputs, to directly extract deep, sharpened spectral features that are robust to motion disturbances locally.

[0080] Specifically, the client loads the data packet passed from step S412, containing the original blurred hyperspectral image data and a precisely matched, calculated set of point spread functions (PSFs). To adapt to the network input size and preserve local context, the original blurred spectral image is spatially divided into overlapping image patches. Simultaneously, each image patch is indexed and its corresponding PSF kernel is extracted.

[0081] The client loads a locally pre-built motion-invariant feature network. This network has a specially designed dual-branch or multimodal input interface. The network's front-end layer receives a blurred spectral image patch and the corresponding PSF kernel as joint input. One encoding branch processes the spectral image patch, and the other encoding branch processes the PSF kernel (treating it as a parameterized condition describing the degradation process). In shallow layers, the network deeply fuses the physical degradation prior information encoded by the PSF with the apparent information of the blurred image through mechanisms such as cross-attention or feature concatenation, forming an intermediate representation that knows how it is blurred.

[0082] The deep layers of the network consist of a series of learnable deconvolutional or feature upsampling layers interwoven with invariant feature extraction layers. Guided by the PSF condition, the deconvolutional layers learn to reverse the blurring process, attempting to recover the image's structure and details in the feature space. However, it does not aim to output a pixel-perfectly sharp image. Invariant feature extraction layers: This is the core of the network. These layers are trained to ignore distortions introduced by various motion patterns described by the PSF, focusing instead on extracting essential spectral-spatial patterns that remain stable even in blurred states and are relevant to the mineral category. The network's optimization objective is directly related to the performance of subsequent classification tasks, forcing it to learn motion-robust features.

[0083] The network ultimately outputs not a reconstructed, sharp image, but a high-dimensional, deep spectral-spatial feature map. Semantically, this feature map is close to the high-level features extracted from a sharp image acquired in an ideal, static state. This feature map possesses distinct characteristics: high feature discrimination between different mineral regions, high consistency of features within the same mineral type, and insensitivity to motion-induced changes in the input data (motion invariance).

[0084] The client spatially reassembles the deep feature maps output by the network for all image patches, forming a full-image deep feature map corresponding to the space of the original input image. This full-image deep feature map is the motion-sensitive, deep-clarification spectral feature map used in this step.

[0085] S414 converts the sharpening features output by the network into enhanced hyperspectral image data.

[0086] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By integrating motion sensing and correction capabilities locally on the client side, the problem of hyperspectral data blurring and degradation caused by the movement and vibration of downhole probes is effectively solved, thus significantly improving the accuracy of mineral identification. Under the premise of strictly protecting the data privacy of each mining area, this solution enables the federated learning system to extract robust features from higher-quality, clearer borehole wall spectral data. This not only enhances the adaptability and generalization ability of the global model in complex dynamic acquisition environments but also lays a solid data foundation for achieving accurate and reliable cross-mining area collaborative identification.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-mining area spectral mineral identification optimization method based on federated learning, characterized in that, include: S1. A global mineral identification model with a neural network architecture for processing hyperspectral data is built in the central server. Each participating mining area, as a client, deploys the same model architecture locally and prepares its own private historical spectral dataset. S2, each client uses a private historical spectral dataset and few-shot augmentation technology to pre-train a model locally and inject domain prior knowledge; S3, each client encrypts the gradient update of the trained model and uploads it to the central server. The server uses a weighted algorithm designed for non-independent and identically distributed models to perform secure aggregation and generate a new generation of global models. S4, each client establishes a local knowledge consolidation library to store key knowledge representations and continuously monitors for newly added local exploration data; S5: When the new data reaches the threshold, the client automatically triggers local training, and combines the new data with the playback of the knowledge consolidation base to update the model using a continuous learning algorithm. In S6, the client uploads the updated gradients in encrypted form, and the server performs secure aggregation again to generate and distribute the next-generation global model after iteration.

2. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 1, characterized in that, The private historical spectral dataset specifically includes: local assets independently owned and controlled by each participating mining area, whose ownership, management rights, and access rights belong to a single mining area entity and are not shared with other participants or a central server; the dataset contains spectral information and corresponding mineral category labels that have been acquired and labeled through historical exploration activities before the start of federated collaborative training; the spectral information in the dataset covers continuous, narrow-band reflectance or emissivity information from visible light to thermal infrared bands; the mineral category labels correspond to each sample in the spectral information and are determined in advance through geological verification methods; the private historical spectral datasets of each participating mining area have significant differences in the spectral characteristics of mineral composition, co-occurrence combinations, and background environment, do not satisfy the independent and identically distributed hypothesis, and there are cases of scarcity of specific mineral category samples and imbalance in the number of samples of different categories.

3. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 1, characterized in that, The next-generation global model includes: a central server collecting privacy-processed model update gradients uploaded by each client; a weighted algorithm aggregating the collected model update gradients to generate an aggregated model update; applying the aggregated model update to the global model parameters of the previous round to generate a next-generation global model; and the central server storing and version-managing the next-generation global model and distributing it to each client.

4. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 1, characterized in that, The continuous learning algorithm updates the model, specifically including: the client continuously monitors newly added exploration data locally, and automatically triggers local training when the accumulated data reaches a preset threshold; during training, the newly added data is mixed with historical core knowledge playback data extracted from the local knowledge consolidation base to form a training set; the continuous learning algorithm is used to train the local model, which prevents the model from forgetting the consolidated historical knowledge when learning new data by introducing regularization constraints or directly replaying historical data; after training is completed, the local knowledge consolidation base is updated, and the model update gradient is calculated.

5. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 1, characterized in that, The continuous monitoring of newly added local exploration data specifically includes: S41, identifying hyperspectral scan data from the borehole wall from the newly added exploration data; S42, performing local synchronous registration of the hyperspectral image and the 3D point cloud for the newly added borehole wall hyperspectral data, and assigning each spectral pixel its corresponding 3D spatial attribute; S43, on the client side, using a physically inspired neural network module, combining 3D geometric features to perform roughness-aware correction on the original borehole wall spectrum, generating a spectrum equivalent to that under standard observation conditions; S44, adopting a spectral-spatial dual-stream network architecture, extracting the deep features of the corrected spectrum and the spatial context features of the 3D point cloud in parallel on the client side, and performing adaptive fusion to complete mineral identification; S45, generating enhanced samples from the corrected high-quality data and identification results, and extracting key anti-interference knowledge from them and storing them in a local knowledge consolidation base; wherein, the 3D spatial attributes include surface normal vector and local curvature, used to quantitatively describe the roughness and micro-topography of the borehole wall after cleaning.

6. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 5, characterized in that, The generation of a spectrum equivalent to that under standard observation conditions includes: locally on the client side, combining the original pore wall spectrum with its three-dimensional geometric features and inputting it into a physics-inspired neural network correction module; the network learns the mapping relationship from actual rough observation conditions to standard flat observation conditions by integrating prior physical laws of the interaction between light and rough surfaces; the network outputs a spectral correction value and uses this correction value to process the original pore wall spectrum; after processing, a corrected spectrum is generated, which is equivalent to the spectrum of the target mineral measured under standard laboratory observation conditions.

7. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 5, characterized in that, The mineral identification process includes: employing a spectral-spatial dual-stream network architecture to process the spectral information and three-dimensional spatial information in parallel on the client side; extracting deep spectral features and spatial context features through the dual-stream network; dynamically fusing the spectral features and spatial features through an adaptive fusion module in the multi-layer structure of the network; and inputting the features generated after the final hierarchical fusion into a classifier to output the mineral category and its confidence level assessment for each spatial unit.

8. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 5, characterized in that, The process of identifying hyperspectral scanning data from the borehole wall in the newly added exploration data is as follows: S411, after identifying the hyperspectral data from the borehole wall, simultaneously check whether it carries complete motion data of the inertial measurement unit, and automatically trigger or skip the motion fuzz correction process accordingly. S412, parsing motion data locally on the client side, and calculating and reconstructing the point spread function that causes spectral image degradation by combining the scanning model; S413, using a specially designed motion-invariant feature network, taking the original blurred spectrum and point spread function as joint inputs, directly extracting deep-level sharpening spectral features that are robust to motion interference locally. S414 converts the sharpening features output by the network into enhanced hyperspectral image data.

9. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 8, characterized in that, The automatic triggering or skipping of motion blur correction process includes: the client checking whether the newly added aperture wall hyperspectral data carries complete inertial measurement unit motion data, and performing integrity checks on the motion data for timestamp alignment, data dimension completeness, and sampling rate coverage; if the motion data passes the integrity check, the subsequent motion blur correction process is automatically triggered; if the motion data does not carry the data or the check fails, the motion blur correction process is skipped.

10. The cross-mining area spectral mineral identification optimization method based on federated learning according to claim 8, characterized in that, The deep-clarification spectral features specifically include: inputting the blurred spectral information and the corresponding point spread function as joint inputs into a motion-invariant feature network; the network reverses the blurring process in the feature space by fusing the physical degradation prior encoded by the point spread function; the network focuses on extracting essential spectral-spatial patterns that are robust to motion disturbances and related to mineral categories through an invariant feature extraction layer; and finally outputs a deep spectral-spatial feature map, which is semantically equivalent to the features extracted from a still, clear image.

Images